The Tissue Moment for Large Language Models

People used to blow their noses with handkerchiefs.

A handkerchief could last for years—washed when dirty, dried, and reused.

Then tissues arrived.

What truly changed people wasn’t that tissues were far superior to handkerchiefs, but that their cost dropped to a critical point:

Blow once, toss it away.

No one would wash a tissue, hang it to dry, and use it again the next day.

Because when something becomes cheap enough, people don’t just “use a little more”—they fundamentally change how they use it.

I think large language models are approaching their own such moment.

I call it:

The Tissue Moment for Large Language Models.

Today’s AI Is Still a Handkerchief

Despite talk of Tokens becoming cheaper, we’re still quite frugal with how we use large language models.

We compress Prompts.

Cache Contexts.

Reduce model calls.

Hope an Agent gets things right in one go.

Even our current vision of Agents resembles hiring a long-term employee:

Give it memory, skills, and let it accumulate experience for repeated use.

Because intelligence today still has costs.

Not just money, but time.

One Agent thinking for 30 seconds is fine.

100 Agents thinking for 30 seconds each starts to feel extravagant.

But what if one day, Tokens become fast and cheap enough?

We’ll Start “Wasting Intelligence”

Then, a task might no longer go to a single Agent.

Instead, we might directly generate 100.

10 to research the problem.

20 to propose solutions.

30 to execute directly.

20 to specifically find flaws.

The rest to validate and vote.

Finally, one answer remains.

The other 99 Agents?

Discarded.

This sounds wasteful.

But when tissues first appeared, “blow once and discard” probably seemed wasteful too.

The real change lies here:

When the marginal cost of intelligence is low enough, humans will shift from “reusing intelligence” to “consuming intelligence.”

The true productivity leap in the AI era may not be AI finally making no mistakes.

Instead:

We can finally afford to let AI make mistakes at scale.

An Agent fails once? Try again.

99 out of 100 proposals are wrong? No problem.

If failure is cheap enough, trial and error itself becomes industrialized.

Agents Might Not Be Employees—They’re Consumables

I’m increasingly skeptical of the popular saying:

Agents will become digital employees.

Maybe some will.

But many others may be more like disposable consumables.

For example, researching a company.

The system temporarily generates a financial Agent, an industry Agent, a bearish Agent, a bullish Agent, and a validation Agent.

Task complete—all destroyed.

Next time, regenerate from scratch.

They don’t even need names, long-term memory, or “growth.”

Born for the task.

Receive the context.

Complete the task.

Destroyed.

Very much like ephemeral instances in cloud computing.

One Step Further: Software Could Become Tissues Too

Why was software reused in the past?

Because it was expensive.

No one would develop an entire app just to process one Excel sheet.

So for decades, software logic has been:

Build once, use many times.

Develop once, serve as many users as possible.

But what if generating software itself becomes cheap enough?

You receive an Excel sheet.

AI generates a small, custom app just for that sheet—for filtering, charts, anomaly analysis, and export—tailored to this single task.

You use it for 20 minutes.

Close it.

That software never appears again.

Next sheet, generate a new one.

This becomes:

Generate once, use once.

Software shifts from asset to consumable.

Agents shift from employees to disposable labor.

Reasoning shifts from a scarce resource to a wasted resource.

Truly Scarce Things Will Also Change

Of course, cheap Tokens don’t mean everything is free.

Models may generate 100 proposals in a second, but who tells them which is correct?

Code still needs to run.

Experiments still need to be conducted.

Users still need to pay.

Robots still need to move in the real world.

So as “thinking” becomes cheaper, what’s truly expensive may become:

Feedback and validation.

Once models are abundant in ideas, the scarcest resources become:

Real environments.

Real data.

Real users.

Real execution outcomes.

And a mechanism to judge “what’s actually correct.”

So the strongest Agents of the future might not just be the smartest models.

But the systems best at:

Generate → Execute → Validate → Discard → Regenerate.

Today we still seriously discuss:

Whether a Prompt can be reused.

How much more we can trim the Context.

Why this task called the model 200 times.

But perhaps years from now, we’ll look back and think it’s a bit like:

A person standing in a supermarket, seriously studying whether a tissue used to blow their nose can be washed, dried, and reused the next day.

When technology truly changes the world, it’s rarely because something becomes 10% cheaper.

It’s when it becomes cheap enough that:

We suddenly become willing to waste it.

Handkerchiefs didn’t disappear because tissues appeared.

Long-term software won’t disappear.

Long-term Agents won’t disappear.

But another world might be emerging:

Disposable Tokens.

Disposable Agents.

Disposable software.

Disposable intelligent labor.

Use and discard.

This might be the true tissue moment for large language models.